Definition
Plain language
A bit of text tacked onto a model's transcript that pushes it to commit to a final answer right now.
As stated in the literature
In chain-of-thought interpretability, an appended cue (e.g., an end-of-thinking marker plus 'therefore the answer is') used to make a model emit its current best answer at a truncation point, measuring when the answer stabilizes; the forcing is itself an intervention that biases toward commitment, which weakens claims that the answer was 'already fixed.'
Why it matters: It reveals when a model's answer stabilizes during reasoning, though the nudge itself can push the model toward committing early and muddy the conclusion.
For example, a researcher might cut off a model mid-reasoning and add 'therefore the answer is' to see what answer it commits to at that exact point.
Heard on the show
“It's the answer-forcing suffix.”Episode 140 — When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided